Ensemble Machine Learning And Deep Learning Framework For Flood Susceptibility Mapping In The Transboundary Rapti River Basin

Flood susceptibility mapping is crucial to reduce the effects of frequent floods, particularly in data-poor and transboundary catchments. In this research, an innovation probability-calibrated hybrid ensemble approach combining Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) networks was used to map the flood-prone areas in the Indo–Nepal Rapti River Basin. In contrast to traditional single-model strategies, the new method simultaneously models spatial heterogeneity and temporal flood behavior, includes uncertainty quantification, and uses spatially blocked cross-validation for improved generalization. The methodology is based solely on open-access data—Sentinel-1 SAR, Sentinel-2 optical images, SRTM DEM, and Copernicus LULC—and eight flood-conditioning variables (elevation, slope, Stream Power Index, Topographic Wetness Index, drainage density, distance to river, NDVI, and land use/land cover). 2017 and 2021 historical floods based on SAR imagery and compared with Central Water Commission bulletins and NASA GFMS were employed for training and validation. The hybrid ensemble exhibited the best predictive ability (AUC = 0.85) with 15–20% improvement in recall of flooded classes over individual models and more than 85% spatial match with historical flood areas. The method can be applied to other monsoon-induced, data-scarce, and transboundary basins and can provide an economical and scalable solution for operational flood risk zoning, climate adaptation planning, and early warning systems.